Leveraging LLMs to identify patent-free glass materials

S Urata and N Miya and R Hashimoto and T Konishi and K Fujii and Y Murakami and K Kanehara, COMPUTATIONAL MATERIALS SCIENCE, 258, 114022 (2025).

DOI: 10.1016/j.commatsci.2025.114022

Materials informatics (MI), enhanced with machine-learning technologies, allows us to efficiently explore a wide range of materials based on experimental and theoretical databases. However, even if materials are found to be superior in some properties, they are often hindered from entering the market due to patent restrictions. Therefore, it is essential to identify a group of patent-free materials before launching a research project in the industry. Nevertheless, the labor-intensive task of identifying the coverage of patented materials by referring to numerous IP documents hinders the application of knowledge in materials discovery. In this work, we utilized a large-language model (GPT-4) to construct a database (IP-DB), which captures composition ranges claimed in approximately 900 granted and issued patents related to high- refractive glass materials. We extracted entities such as ingredients, their minimum and maximum amounts, units, and inequalities from IP documents using few-shot prompts. Using the IP-DB, possible candidates satisfying target properties of density, refractive index, and liquidus temperature were identified to be patent-free. To estimate these properties, we applied a machine-learning model, GlassNet. Simultaneously, density functional theory (DFT) calculations on the glass structures elaborated by a machine-learning simulator were used to theoretically evaluate refractive indices to confirm the reliability of those estimated by the empirical model. Consequently, the workflow combining the LLM, MI, machine-learning simulator, and DFT calculations found compositions worth investigating to develop high-refractive glass applications. In contrast, the candidate compositions did not form transparent glass, indicating the necessity to develop a reliable method for estimating glass-forming ability of materials.

Return to Publications page